The AI Cheat Sheet.

R-T-C-F · The universal prompt formula

Role · Task · Context · Format. Works for 80% of prompts.

You are a [ROLE with specific expertise].
Your task is to [SPECIFIC OUTCOME, not a topic].
Context you need: [audience, constraints, brand voice, examples].
Return the answer as [FORMAT: numbered list / table / 200-word email / JSON].

C-G-S-R · Strategy prompts

Context · Goal · Steps · Risks.

Context: [the situation in 2-3 lines].
Goal: [the single outcome I want].
Walk me through the steps you would take, in order.
End with the top 3 risks and how to mitigate each.

P-A-S · Copywriting (Problem-Agitate-Solve)

For ads, emails, landing pages.

Write [DELIVERABLE] for [AUDIENCE] selling [PRODUCT].
Use the Problem-Agitate-Solve structure:
1. Name their problem in their own words.
2. Agitate the cost of not solving it.
3. Reveal [PRODUCT] as the obvious solution.
Tone: [conversational / direct / premium]. Length: [N words].

Self-Critique Loop

Make the model fix its own output before you see it.

First, produce a draft of [DELIVERABLE].
Then, critique your draft against these criteria: [criteria].
Then, rewrite the draft incorporating every critique.
Return only the final, rewritten version.

Ladder of Specificity

Push vague requests into useful ones.

I want to write [VAGUE GOAL].
Before you write anything, ask me 5 questions that will most change the quality of the final output.
Wait for my answers, then produce the deliverable.

Teach-Back

For learning a new concept fast.

Explain [CONCEPT] to me in 3 layers:
1. A one-sentence definition a 12-year-old understands.
2. A 200-word explanation for a smart professional.
3. The 3 nuances most beginners get wrong.
Then quiz me with 3 questions.

Decision Matrix

Stop ruminating. Force a structured choice.

I'm deciding between [OPTION A] and [OPTION B] for [GOAL].
Build a 5-row decision matrix scoring each on: cost, time, reversibility, upside, risk.
Score 1-5, sum the totals, and tell me which to pick and why.

Deep Research Brief

For research before a meeting/pitch.

Give me a 1-page brief on [COMPANY / PERSON / TOPIC].
Cover: what they do, who their customers are, recent moves, public positioning, likely priorities right now, and 3 angles I could open a conversation with.
Cite sources. If unsure, say so.

Few-shot examples

Show 2-3 input/output pairs. Quality jumps immediately.

Here are examples of the style I want:

Input: [example 1 input]
Output: [example 1 output]

Input: [example 2 input]
Output: [example 2 output]

Now do the same for:
Input: [your real input]
Output:

Chain-of-thought

Force the model to reason before answering.

Think step-by-step. Show your reasoning before the final answer. Then put the final answer on a new line prefixed with "ANSWER:".

ReAct (Reason + Act)

For agents using tools.

For each step, output:
Thought: [why you're doing this]
Action: [the tool / call to make]
Observation: [the result]
Repeat until done, then output Final Answer.

Stepback prompting

Solve the abstract problem first, then the concrete one.

Before answering [SPECIFIC QUESTION], first answer the more general question it sits inside. Then use that to answer the specific one.

Persona priming

Lock the model into a role for the whole chat.

For this entire conversation you are [NAMED PERSONA].
You have [N years of experience] in [DOMAIN].
You speak [tone].
You never [forbidden behaviour].
Confirm you understand before I start.

Negative constraints

Telling it what NOT to do is often stronger than what to do.

Constraints:
- Do NOT use [overused word / cliché].
- Do NOT start sentences with "In today's...".
- Do NOT add disclaimers or apologies.
- Do NOT exceed [N] words.

ChatGPT (GPT-5 / 4o)

Best all-rounder for chat, writing, and image generation in one tool.

Strengths: creative writing, image generation, voice mode, real-time browse.
Weakness: less rigorous reasoning than Claude on long technical work.

• •Pick for: marketing, brainstorming, fast drafts, image gen
• •Context: ~128k tokens
• •Pricing tier: $20/mo Plus, $200/mo Pro

Claude (Sonnet 4.5 / Opus)

Best for long reasoning, writing tone, and coding.

Strengths: nuanced writing, large context handling, careful reasoning, code.
Weakness: no native image generation, can be over-cautious.

• •Pick for: long docs, coding, strategy, careful writing
• •Context: 200k tokens
• •Pricing tier: $20/mo Pro, $100/mo Max

Gemini (2.5 Pro)

Best for huge context and Google Workspace integration.

Strengths: 1M-token context, multimodal video/audio, Workspace integration.
Weakness: writing voice is the weakest of the big 3.

• •Pick for: ingesting whole codebases / documents
• •Context: up to 1M tokens
• •Pricing tier: $20/mo AI Pro

Grok (3)

Best for real-time X/Twitter context.

Strengths: live web + X search, less restrictive guardrails.
Weakness: writing quality and reliability still trail GPT/Claude.

• •Pick for: real-time research, niche communities
• •Context: 128k tokens

Llama 4 (open-source)

Best for self-hosted / private workloads.

Strengths: free, deployable on your own infra, no data leaves your stack.
Weakness: needs GPU; quality below top closed models.

• •Pick for: privacy-sensitive automations
• •Run via: Ollama / Groq / Together

Quick model picker

Cheat sheet: 'I want to X → use Y'.

Write social copy / brainstorm / make an image → ChatGPT
Write a long proposal / a careful email / code → Claude
Analyse a 300-page PDF or a whole codebase → Gemini
Real-time news, X-trends, niche research → Grok / Perplexity
Anything that must stay on your servers → Llama (self-hosted)

Temperature

Controls creativity vs. consistency.

0.0–0.3 → factual, deterministic (data, code, SQL, classification)
0.4–0.7 → balanced (most writing tasks)
0.8–1.2 → creative, varied (brainstorming, taglines, fiction)

Top-p (nucleus sampling)

Alternative way to control variety. Don't tune both.

Default 1.0 in most apps. Lower to 0.7–0.9 if you want less repetition without changing temperature.

System vs user messages

System sets the rules. User sends the task.

System: persistent persona, format rules, what it must never do.
User: the current request.
Putting persona in the user message is the #1 reason ChatGPT 'forgets' instructions mid-chat.

Tokens, in plain English

1 token ≈ ¾ of an English word. A page ≈ 500 tokens.

Why it matters:
- Context windows are in tokens (128k = ~96k words).
- API pricing is per token.
- Long outputs cost more than long inputs.
Rule: front-load instructions, push examples and reference docs to the end.

Vague verbs ('write something good')

AI mirrors your vagueness back at you.

Replace 'write something good' with 'write a 120-word LinkedIn post in [voice] aimed at [audience] driving them to [action]'.

Asking for tone without an example

Style transfer is far more accurate than style description.

Paste 2-3 paragraphs of writing you love. Tell it: 'Match this voice exactly'. Then give the task.

Asking for 5 unrelated things in one prompt

Quality collapses past 2-3 sub-tasks.

Split into separate chats or break into a numbered, sequential prompt: 'Step 1 ... wait. Step 2 ... wait.'

Trusting numbers, citations, and quotes

LLMs invent these confidently. Always verify.

Rule: any number, date, citation, legal claim, or quote it generates must be verified before you publish or send.

Using one giant chat for everything

The longer the chat, the more it drifts and forgets.

Start a new chat per project. Save your reusable system prompt. Paste relevant context in fresh — don't expect it to remember last week's chat.

1 idea → 7 pieces of content

Repurpose any insight into a week of content.

Take [SOURCE: blog / podcast / customer call].
Produce:
1. One Instagram carousel (7 slides, headline + body each)
2. One LinkedIn post (180 words, hook first)
3. One X/Twitter thread (8 tweets)
4. One YouTube short script (45 seconds)
5. One email to my list (250 words)
6. One TikTok hook + 30-sec script
7. One blog post outline (H1 + 5 H2s)
Keep voice consistent. Adapt format, not message.

Cold outreach in 3 steps

Personalised cold DM / email at scale.

Step 1: Summarise [LEAD'S public profile / company site] in 3 bullets — what they do, who they serve, what's changed recently.
Step 2: Find one specific hook from step 1 I could open with.
Step 3: Draft a 4-line DM: opener referencing the hook, one line on what I do, one line on the value, soft CTA. No fluff, no emojis.

Meeting → next actions

Turn raw notes into clear deliverables.

Here are my raw meeting notes: [PASTE].
Produce:
1. A 5-bullet summary
2. Decisions made (with owner if mentioned)
3. Action items (owner + deadline)
4. Open questions still to resolve
5. A 3-line follow-up email I can send to attendees

Launch a product in 1 prompt

Full launch kit from a product description.

Product: [DESCRIPTION].
Audience: [WHO].
Generate a launch kit:
- One landing page outline (hero, 3 features, social proof, FAQ, CTA)
- 3 email sequence (announce, value, last-call)
- 5 social posts (mix of teaser, value, proof, story, hard CTA)
- 1 launch-day script for a 60-sec video
Match the tone of: [PASTE 1 paragraph you like].